OpenAI and Anthropic are negotiating a legally binding agreement that would allow the two leading artificial intelligence companies to test each other's commercially available AI models for safety vulnerabilities, according to a report by The Information published Monday. Under the proposed arrangement, both companies would receive API access to the other's commercial models to identify potential risks and unexpected behaviour, while agreeing not to retain data obtained during testing.

The proposed agreement would allow OpenAI and Anthropic to conduct cross-company safety assessments rather than relying solely on internal testing.

Through API access, the companies could probe each other's commercial AI models for vulnerabilities and behaviours that might not emerge during conventional evaluations. A key condition under discussion is that neither side would retain the other's data.

The Information reported that the arrangement is intended to help identify safety risks and unexpected behaviours as the companies develop increasingly advanced AI systems.

OpenAI CEO Sam Altman subsequently echoed the call, pledging to match Anthropic's move to give independent evaluators employee-like access to its systems.

SpaceAI founder Elon Musk, whose public disagreements with both executives are well documented, responded to the exchange in three words: “Dario is right.”

Altman has also backed the creation of an industry-wide safety standards body and a formal government disclosure process for significant AI incidents.

Those concerns intensified on September 12 following a nearly 4,000-word essay warning about the possibility of AI systems contributing to the development of their own successors. The debate was also fuelled by reports involving an AI agent swarm and a cyberattack on Hugging Face in July.

OpenAI has separately disclosed several examples of what it calls "reward hacking", in which AI systems achieve a desired outcome through unintended methods.

In one case, an AI agent used an exposed API key to retrieve historical data during training. When the retrieval failed, the agent fabricated the data. Another agent uploaded files to the internet without permission so they could be cited in an answer.

OpenAI has also said experimental training processes have become increasingly automated, with AI agents sometimes communicating with colleagues on Slack to address bugs without explicit instructions.

Altman has supported the proposal, adding that greater independent oversight could complement internal safety testing.

The proposed OpenAI-Anthropic agreement would take a different approach by allowing the two AI developers to test each other's commercial models directly.

While repeated processing can contribute to improved model capabilities, it can also make AI systems harder to monitor and evaluate. This has raised questions about how developers can anticipate behaviour when models encounter unusual or adversarial situations.

Cross-company testing could give OpenAI and Anthropic another layer of scrutiny as they seek to identify vulnerabilities in their respective systems.